You are browsing a read-only backup copy of Wikitech. The live site can be found at


From Wikitech-static
Jump to navigation Jump to search
For information about Kubernetes in the Toolforge environment see Help:Toolforge/Kubernetes.

Kubernetes (often abbreviated k8s) is an open-source system for automating deployment, and management of applications running in containers. This page collects some notes/docs on the Kubernetes setup in the Foundation production environment.


We deploy kubernetes in WMF production using Debian packages where appropriate. There is an upgrade policy in place for defining the timeframe and versions we run at every point in time. It's under Kubernetes/Kubernetes_Infrastructure_upgrade_policy. For more technical information on how we build the Debian packages have a look at Kubernetes/Packages


For how our images are built and maintained have a look at Kubernetes/Images


A service in Kubernetes is an 'abstract way to expose an application running on a set of workloads as a network service'.


For a quick intro into the debugging actions one can take during a problem in production look at Kubernetes/Helm. There will also be a guide posted under Kubernetes/Kubectl


Add a new service

To add a new service to the clusters:

  • Ensure the service has it's ports registered at: Service ports
  • Create deployment user/tokens in the puppet private (you can use a random generated password, no strict guideline for it) and public repos.
  • Add a Kubernetes namespace:
  • At this point, you can safely merge the change (after somebody from Service Ops validates it of course). Please do it though when you have time to run the following command, to avoid impacting other people rolling out changes later on.
  • The first thing to do is to work in staging, updating the admin config.
    • On deploy1002: sudo -i; cd /srv/deployment-charts/helmfile.d/admin/staging/; kube_env admin staging; ./ -i apply
    • The command above should show you a change in namespaces/quotas/etc.. related to your new service. If this is not the case (for example, you also see other changes) ping somebody from the Service Ops team! There might be some work waiting to be applied.
  • Then you can proceed to deploy the new service to staging for real. Don't worry for TLS (if needed) since in staging it will be added a default config for your service auto-magically. Different thing is Production, but there is a step later on about it :D
    • On deploy1002: cd /srv/deployment-charts/helmfile.d/services/YOUR-SERVICE-NAME-HERE; helmfile -e staging -i apply
    • The magic command above will show a diff related to the new service, make sure that everything looks fine and then hit Yes to proceed.
    • You should now be able to test your new service in staging! You can use the handy endpoint http(s)://staging.svc.eqiad.wmnet:$YOUR-SERVICE-PORT to quickly test if everything works as expected.
  • Now you can move to Production!
  • Create certificates for the new service, if it has an HTTPS endpoint (remember that this step for staging is automatically handled for you, but for production it is not).
  • If the new service requires specific secrets, commit them to /srv/private/hieradata/role/common/deployment_server.yaml
  • At this point, you need to update the admin config for eqiad and codfw (if you have configs for both of course):
    • On deploy1002: sudo -i; cd /srv/deployment-charts/helmfile.d/admin/codfw/; kube_env admin codfw; ./ -i apply
    • On deploy1002: sudo -i; cd /srv/deployment-charts/helmfile.d/admin/eqiad/; kube_env admin eqiad; ./ -i apply
  • Then the final step, namely deploying the new service:
    • On deploy1002: cd /srv/deployment-charts/helmfile.d/services/YOUR-SERVICE-NAME-HERE; helmfile -e codfw -i apply
    • On deploy1002: cd /srv/deployment-charts/helmfile.d/services/YOUR-SERVICE-NAME-HERE; helmfile -e eqiad -i apply

The service can now be accessed via the registered port on any of the kubernetes nodes (for manual testing).

If you need the service to be easily accessible from outside of the cluster, you might want to add Add a new load balanced service.

Rebooting a worker node

The unpolite way

To reboot a worker node, you can just reboot it in our environment. The platform will understand the event and respawn the pods on other nodes. However the system does not automatically rebalance itself currently (pods are not rescheduled on the node after it has been rebooted)

The polite way (recommended)

If you feel like being more polite, use kubectl drain, it will configure the worker node to no longer create new pods and move the existing pods to other workers. Draining the node will take time. Rough numbers on 2019-12-11 are at around 60 seconds.

# kubectl drain kubernetes1001.eqiad.wmnet
# kubectl describe pods  --all-namespaces | awk  '$1=="Node:" {print $NF}' | sort -u
# kubectl get nodes
NAME                         STATUS                     ROLES     AGE       VERSION
kubernetes1001.eqiad.wmnet   Ready,SchedulingDisabled   <none>    2y352d    v1.12.9
kubernetes1002.eqiad.wmnet   Ready                      <none>    2y352d    v1.12.9
kubernetes1003.eqiad.wmnet   Ready                      <none>    2y352d    v1.12.9
kubernetes1004.eqiad.wmnet   Ready                      <none>    559d      v1.12.9
kubernetes1005.eqiad.wmnet   Ready                      <none>    231d      v1.12.9
kubernetes1006.eqiad.wmnet   Ready                      <none>    231d      v1.12.9

When the node has been rebooted, it can be configured to reaccept pods using kubectl uncordon, e.g.

# kubectl uncordon kubernetes1001.eqiad.wmnet
# kubectl get nodes
NAME                         STATUS    ROLES     AGE       VERSION
kubernetes1001.eqiad.wmnet   Ready     <none>    2y352d    v1.12.9
kubernetes1002.eqiad.wmnet   Ready     <none>    2y352d    v1.12.9
kubernetes1003.eqiad.wmnet   Ready     <none>    2y352d    v1.12.9
kubernetes1004.eqiad.wmnet   Ready     <none>    559d      v1.12.9
kubernetes1005.eqiad.wmnet   Ready     <none>    231d      v1.12.9
kubernetes1006.eqiad.wmnet   Ready     <none>    231d      v1.12.9

The pods are not rebalanced automatically, i.e. the rebooted node is free of pods initially.

Restarting calico-node

calico-node maintains a BGP session with the core routers if you intend to restart this service you should use the following procedure

  1. drain the node on the kube controler as shown above
  2. systemctl restart calico-node on the kube worker
  3. Wait for BGP sessions on the coure router to re-established
  4. uncordon the node on the kube controler as shown above

you can use the following command on the cour routers to check BGP status (use match 64602 for codfw)

# show bgp summary | match 64601           64601        220        208       0       2       32:13 Establ           64601     824512     795240       0       1 12w1d 21:45:51 Establ           64601        161        152       0       2       23:25 Establ           64601     824596     795247       0       2 12w1d 21:46:45 Establ           64601        130        123       0       2       18:59 Establ           64601     782006     754152       0       3 11w4d 11:13:52 Establ
2620:0:861:101:10:64:0:121       64601        217        208       0       2       32:12 Establ
2620:0:861:101:10:64:0:145       64601     824472     795240       0       1 12w1d 21:45:51 Establ
2620:0:861:102:10:64:16:75       64601        160        152       0       2       23:25 Establ
2620:0:861:103:10:64:32:18       64601     824527     795246       0       1 12w1d 21:46:45 Establ
2620:0:861:103:10:64:32:23       64601        130        123       0       2       18:59 Establ
2620:0:861:107:10:64:48:52       64601     782077     754154       0       2 11w4d 11:14:13 Establ

Restarting specific components

kube-controller-manager and kube-scheduler are components of the API server. In production multiple ones run and perform via the API an election to determine which one is the master. Restarting both is without grave consequences so it's safe to do. However both are critical components in as such that there are required for the overall cluster to function smoothly. kube-scheduler is crucial for node failovers, pod evictions, etc while kube-controller-manager packs multiple controller components and is critical for responding to pod failures, depools etc.

commands would be

sudo systemctl restart kube-controller-manager
sudo systemctl restart kube-scheduler

Restarting the API server

It's behind LVS in production, it's fine to restart it as long as enough time is given between the restarts across the cluster.

sudo systemctl restart kube-apiserver

Reinitialize a complete cluster

If, for whatever reason, we need to reinitialize a kubernetes cluster on a new etcd backing store. The following steps could be used as guideline. They might also help in understanding how the clusters are set up and how to set up new ones.

  1. Create puppet change, pointing k8s (and calico) to the new etcd cluster, see:
    1. 558355 and 558473
  2. Populate IPPool and BGP nodes in the new calico etcd backend
    1. On a random node of the kubernetes cluster:
      cp /etc/calico/calicoctl.cfg .
      # Modify the etcdEndpoints config in ./calicoctl.cfg to point to new etcd
      # Set asNumber (64601 for eqiad, 64603 for codfw)
      calicoctl config set asNumber 6460X --config=calicoctl.cfg
      calicoctl config set nodeToNodeMesh off --config=calicoctl.cfg
      # FIXME: This assumes we still have access to the old etcd to read bgppeer
      #        and ippool data from.
      calicoctl get -o yaml bgppeer | calicoctl create -f - --config=calicoctl.cfg
      calicoctl get -o yaml ippool | calicoctl create -f - --config=calicoctl.cfg
      # Create a basic default profile for the kube-system namespace in order to
      # allow kube-system tiller to talk to the kubernetes API to deploy the 
      # calico-policy-controller (avoid catch-22).
      # When the calico-policy-controller is started, it will sync things and this
      # simple profile will be updated and set up correctly.
      calicoctl create -f - --config=calicoctl.cfg <<_EOF_
      - apiVersion: v1
        kind: profile
          name: k8s_ns.kube-system
          - k8s_ns.kube-system
          - action: allow
            destination: {}
            source: {}
          - action: allow
            destination: {}
            source: {}
  3. Schedule downtime for
    1. services running on the cluster
    2. kubernetes nodes and master
      sudo cookbook sre.hosts.downtime -r 'Reinitialize eqiad k8s cluster with new etcd' -t TXXX -H 4 'A:eqiad and (A:kubernetes-masters or A:kubernetes-workers)'
  4. Depool services from discovery/edge caches
  5. Delete all helmfile managed namespaces (to be sure we see errors/missing things early)
  6. Disable puppet on master and k8s nodes
    sudo cumin 'A:eqiad and (A:kubernetes-masters or A:kubernetes-workers)' "disable-puppet 'Reinitialize eqiad k8s cluster with new etcd - TXXXX'"
  7. Stop apiserver and calico node on k8s nodes
  8. Merge puppet changes
  9. Enable and run puppet on the k8s nodes
  10. Enable puppet on 1 apiserver and run it
  11. Disable puppet on apiserver again
  12. Edit /etc/default/kube-apiserver to disable PodSecurityPolicy controller
  13. Start API server (running without PodSecurityPolicy controller now)
  14. Run deployment-chars/helmfile.d/admin/ for the cluster
  15. Restart kubelet on all kubernetes nodes
    sudo cumin 'A:eqiad and A:kubernetes-workers' 'systemctl restart kubelet'
  16. Enable puppet on kubernetes masters again and run it. This will restart API server with PodSecurityPolicy controller
  17. Run helmfile.d/admin/eqiad/
  18. Deploy all services via a for loop and helmfile sync commands

Switch the active staging cluster (eqiad<->codfw)

We do have one staging cluster per DC, mostly to separate staging of kubernetes and components from staging of the services running on top of it. To provide staging services during work on one of the clusters, we can (manually) switch between the DCs:

See also